Modelle

0GM-1.0-35B-A3B

0G.AI in-house model optimized for agentic coding and tool use; thinking enabled by default. Stream long outputs: non-streaming requests must complete within about 10 minutes.

0GM-1.0-35B-A3B-SIA

A 35B hybrid MoE model enhanced with per-token reward guidance. At each decoding step, a 4B Value Model scores candidate tokens and steers generation toward higher-quality outputs, with improvements in harmlessness, helpfulness, and honesty.

Qwen3 Embedding 0.6B

Qwen3-Embedding-0.6B (open weights, Apache-2.0). OpenAI Embeddings API compatible (POST /embeddings: model/input/encoding_format); 1024-dimension dense vectors, 100+ languages including programming languages, 32K-token input. The small, fast member of the family. Instruction-aware: prefix queries with "Instruct: <task>\nQuery: " (documents need no prefix). Billed on input tokens only.

Qwen3 Embedding 8B

Qwen3-Embedding-8B (open weights, Apache-2.0). OpenAI Embeddings API compatible (POST /embeddings: model/input/encoding_format); 4096-dimension dense vectors, 100+ languages including programming languages, 32K-token input. Instruction-aware: prefix queries with "Instruct: <task>\nQuery: " for best retrieval quality (documents need no prefix). Billed on input tokens only.

Vela 2.0 9B

Vela 2.0 9B (vllm-sr, open weights, Apache-2.0): a System One decision model for routing and guardrails, with Jev's request and response shape (POST /decisions: model/state/questions). Answers noul / choice / score questions with calibrated probabilities instead of text, plus set questions (every label above threshold) and span questions (labelled character offsets into the state, e.g. personal information or claims a supplied source does not support). Qwen3.5-based decoder (7.9B), 16,384-token input. Billed on input tokens only.

Zhipu AI GLM-5.3 for long-horizon reasoning, coding, and agentic tool use; 1M context, up to 131K output. Deep thinking supported; whether it is on by default and whether it can be turned off varies by provider. Function calling, JSON mode, and implicit prompt caching supported. Stream long outputs: non-streaming requests must complete within about 10 minutes.

Laya

Laya (Convai Innovations, open weights, Apache-2.0): a non-autoregressive System One decision model, the open alternative to TypeSafe Jev. Same request and response shape as Jev (POST /decisions: model/state/questions, typed noul / choice / score answers with calibrated probabilities instead of text). Routes by language inside the model: English state to a ModernBERT-large checkpoint (512 tokens), other languages to a 100+-language mmBERT checkpoint (1024 tokens); longer state is truncated. Best with fewer than 20 options per choice question. Billed on input tokens only.

Z-Image-Turbo

Asynchronous text-to-image model with Base64 output. Generates at most 2 images per request — requesting more (n > 2) is refused with 400.

Multilingual automatic speech recognition (ASR); transcription and translation.

MiniMax-M3

MiniMax-M3, a natively-multimodal model on MiniMax Sparse Attention (MSA); agentic coding, native tool use, and long-horizon tasks. 1M context, thinking on by default.

Qwen-Flash

Alibaba Qwen-Flash, the cheapest 1M-token-context model of the Qwen line. Text in / text out, hybrid thinking/non-thinking modes, function calling, built-in tools, and structured output.

Qwen3.5-Flash

Alibaba Qwen3.5-Flash, the fast, low-cost tier of the Qwen3.5 series. Text in / text out, 1M-token context, hybrid thinking/non-thinking modes, function calling, and structured output.

Qwen3.7-Flash

Alibaba Qwen3.7-Flash, the fast, low-cost model of the Qwen3.7 series. Text in / text out, 1M-token context, hybrid thinking/non-thinking modes, function calling, built-in tools, structured output, and implicit prompt caching.

Qwen-Flash-Character

Alibaba Qwen-Flash-Character, a role-play / character-dialogue model. Text in / text out, 8K-token context. No function calling and no thinking mode.

Qwen3.5-35B-A3B

Alibaba Qwen3.5-35B-A3B, a 35B-total / 3B-active MoE model of the Qwen3.5 series. Text in / text out, 256K-token context, hybrid thinking/non-thinking modes, function calling, and structured output.

Qwen3.7 Text Embedding

Alibaba Tongyi Lab text embedding model (Qwen3.7-based), OpenAI Embeddings API compatible (POST /embeddings: model/input/dimensions/encoding_format). Single input up to 128,000 tokens and user-selectable output dimension. Billed on input tokens only — the response's `usage` carries no completion_tokens, matching this router's `embedding` service type (input-side billing, no output/completion price).

Qianfan OCR

A document-understanding model for OCR, layout parsing and visual question answering over scanned pages; text and image input, text output, 60K-token context, billed per token rather than per page. Image tokens scale with resolution and are billed through prompt_tokens.

Qwen3.5-27B

Alibaba Qwen3.5-27B, the dense 27B model of the Qwen3.5 series. Text in / text out, 256K-token context, hybrid thinking/non-thinking modes, function calling, and structured output.

Qwen-Plus

Alibaba Qwen-Plus, the long-running balanced model of the Qwen line. Text in / text out, 1M-token context, hybrid thinking/non-thinking modes, function calling, built-in tools, and structured output.

Qwen3.5-122B-A10B

Alibaba Qwen3.5-122B-A10B, a 122B-total / 10B-active MoE model of the Qwen3.5 series. Text in / text out, 256K-token context, hybrid thinking/non-thinking modes, function calling, and structured output.

Qwen3.5-Plus

Alibaba Qwen3.5-Plus, the balanced tier of the Qwen3.5 series. Text in / text out, 1M-token context, hybrid thinking/non-thinking modes, function calling, built-in tools, and structured output.

ERNIE 4.5 Turbo (2026-04-02)

The 2026-04-02 pinned snapshot of Baidu ERNIE 4.5 Turbo, for callers who need behaviour fixed to one build; text in, text out, function calling supported. The upstream does not report cached_tokens, so no cache discount applies.

ERNIE 4.5 Turbo 128K

Baidu ERNIE 4.5 Turbo with a 128K-token context window, a low-cost general-purpose model for chat, extraction and classification; text in, text out, function calling supported. The upstream does not report cached_tokens, so no cache discount applies.

ERNIE 4.5 Turbo 32K

The 32K-token context tier of Baidu ERNIE 4.5 Turbo, priced the same per token as the 128K tier; text in, text out, function calling supported. The upstream does not report cached_tokens, so no cache discount applies.

Zhipu AI GLM-5.3-Flash, a fast, cost-effective model for coding and agentic tasks. Text in / text out, 1M context, up to 128K output. Deep thinking is always on and cannot be disabled. Function calling, JSON mode, and implicit prompt caching supported.

Qwen3-VL-Plus

Alibaba Qwen3-VL-Plus, the high-capability tier of the Qwen3-VL vision-language series: image, video and text in, text out. 256K-token context, hybrid thinking/non-thinking modes, function calling, and structured output.

Qwen3-Coder-Flash

Alibaba Qwen3-Coder-Flash, a fast, low-cost coding model. Text in / text out, 1M-token context, function calling, and structured output.

MiniMax-M3.1-Flash

MiniMax-M3.1-Flash, a natively-multimodal model for reasoning, coding, and agent workflows. Text, image, and video input, text output, 1M-token context, up to 512K output tokens. Thinking is always on and cannot be disabled; reasoning effort has five levels, low through max. Function calling and implicit prompt caching.

Qwen3.8-Flash

Alibaba Qwen3.8-Flash, a fast, cost-effective multimodal model for coding and agentic work; visual and video understanding, function calling, web search, deep thinking, and implicit prompt caching. 1M context.

Qwen3.5-397B-A17B

Alibaba Qwen3.5-397B-A17B, a 397B-total / 17B-active MoE model of the Qwen3.5 series. Text in / text out, 256K-token context, hybrid thinking/non-thinking modes, function calling, and structured output.

Qwen3.6-Flash

Alibaba Qwen3.6-Flash, a fast, low-cost model of the Qwen3.6 series. Text in / text out, 1M-token context, hybrid thinking/non-thinking modes, function calling, built-in tools, and structured output.

Hunyuan-3

Tencent Hunyuan 3 (hy3); 295B total / 21B active MoE, native 256K context. Text in / text out. Function calling and implicit prompt caching supported.

Xiaomi MiMo-V2.6-Flash, the low-cost high-throughput sibling of MiMo-V2.6-Pro. Text + image in / text out, 1M context. Thinking is on by default and can be disabled. Function calling, JSON mode, and implicit prompt caching supported.

ERNIE 5.1

Baidu ERNIE 5.1 for complex reasoning, coding and agentic work; text in, text out, 243K-token context, function calling supported. Input-length tiered pricing: rates rise above 32K input tokens.

Qwen3.6-35B-A3B

Alibaba Qwen3.6-35B-A3B, a 35B-total / 3B-active MoE model of the Qwen3.6 series. Text in / text out, 256K-token context, hybrid thinking/non-thinking modes, function calling, and structured output.

Qwen3-Omni-Flash

Alibaba Qwen3-Omni-Flash, an any-modality-in / text-out model accepting text, image, audio and video input. 48K-token context and function calling.

DeepSeek-V4-Flash-0731

Lightweight MoE (284B total / 13B active) tuned for fast, low-cost, high-throughput text work; function calling, web search, and thinking on by default; 1M context, up to 384K output.

Qwen3.6-Plus

Alibaba Qwen3.6-Plus with hybrid linear attention and sparse MoE; 1M context, 119 languages.

Qwen3.7-Plus

Alibaba multimodal model with vision and video understanding; native function calling, 1M context.

Qwen3.5-Omni-Flash

Alibaba Qwen3.5-Omni-Flash, the fast any-modality-in / text-out model of the Qwen3.5-Omni series, accepting text, image, audio and video input. 192K-token context and function calling.

MiniMax-M2.7

MiniMax-M2.7. Text in / text out, 200K-token context, thinking always on, function calling, and implicit prompt caching.

MiniMax-M2.5

MiniMax-M2.5. Text in / text out, 192K-token context, thinking always on, function calling, and implicit prompt caching.

Qwen-Max

Alibaba Qwen-Max, the top capability tier of the original Qwen line. Text in / text out, 32K-token context, non-thinking only, function calling, built-in tools, and structured output.

Zhipu AI GLM-5.3-FlashX, the higher-throughput tier of GLM-5.3-Flash for coding and agentic tasks. Text + image in / text out, 1M context, up to 128K output. Deep thinking is always on and cannot be disabled. Function calling, JSON mode, and implicit prompt caching supported.

DeepSeek-V4.1-Flash

DeepSeek-V4.1-Flash for fast, low-cost general chat, coding, and agentic work; function calling, structured output, and web search; thinking on by default; 1M context, up to 384K output.

Qwen3.6-27B

Alibaba Qwen3.6-27B, the dense 27B model of the Qwen3.6 series. Text in / text out, 256K-token context, hybrid thinking/non-thinking modes, function calling, and structured output.

ERNIE 4.5 Turbo VL 32K

The 32K-token context tier of Baidu ERNIE 4.5 Turbo VL, priced the same per token as the 128K tier; text and image input, text output.

ERNIE 4.5 Turbo VL

The vision-language build of Baidu ERNIE 4.5 Turbo, for image understanding, visual question answering and document parsing; text and image input, text output, 128K-token context. Image tokens are billed through prompt_tokens at a flat rate per image regardless of resolution.

Xiaomi MiMo-V2.6-Pro, a multimodal reasoning model. Text + image in / text out, 1M context. Thinking is on by default and can be disabled. Function calling, JSON mode, and implicit prompt caching supported. No hidden system prompt is billed.

Xiaomi MiMo-v2.5-Pro. Text in / text out, 1M-token context, thinking on by default and can be disabled, function calling, and implicit prompt caching. A hidden system prompt of roughly 250 tokens is billed as input, so even a one-word request reports about 256 prompt tokens.

DeepSeek-V4-Flash-Vision (experimental)

DeepSeek-V4-Flash-Vision, the experimental vision-enabled build of DeepSeek-V4-Flash. Image and text in, text out, 1M-token context, deep thinking, function calling, and implicit prompt caching.

Moonshot AI Kimi-K2.5. Text in / text out, 256K-token context, hybrid thinking/non-thinking modes, function calling, and implicit prompt caching.

Zhipu AI GLM-5, purpose-built for coding and agent workflows; 744B foundation, 200K context.

Qwen3.8-27B

Alibaba Qwen3.8-27B, the compact dense model of the Qwen3.8 series. Text in / text out, native 1M-token context, hybrid thinking/non-thinking modes, function calling, structured output, and implicit prompt caching.

DeepSeek-V4-Pro-0813

DeepSeek-V4-Pro for agentic coding, multi-step workflows, and complex reasoning; 1M context, up to 384K output.

Zhipu AI GLM-5.2, open-source, purpose-built for long-horizon tasks; 1M lossless context. Strong coding and engineering: autonomous task decomposition, architecture design, full-stack development, integration testing, and multi-platform deployment.

Qwen3.5-Omni-Plus

Alibaba Qwen3.5-Omni-Plus, the high-capability any-modality-in / text-out model of the Qwen3.5-Omni series, accepting text, image, audio and video input. 192K-token context and function calling.

ERNIE 5.0

Baidu ERNIE 5.0, a multimodal model for reasoning over text and images; text and image input, text output, 243K-token context, function calling supported. Input-length tiered pricing: rates rise above 32K input tokens.

ERNIE 5.0 Thinking (preview)

The extended-reasoning variant of Baidu ERNIE 5.0, which returns its chain of thought alongside the answer; text and image input, text output, 243K-token context, function calling supported. Preview: behaviour may change without a new model id.

Moonshot AI Kimi-K2.6. Text in / text out, 256K-token context, hybrid thinking/non-thinking modes, and function calling.

Moonshot AI coding model for agentic coding and tool use; multimodal input (text, image, video), thinking always on. 256K context.

Zhipu AI GLM-5-Turbo, the low-latency turbo variant of GLM-5. Text in / text out, 195K-token context, hybrid reasoning with a deep-thinking mode, function calling, structured output, and implicit prompt caching.

Hunyuan-4-preview

Tencent Hunyuan Hy4 preview; 770B total / 49B active MoE tuned for agent, coding and production workflows — stronger task decomposition, long-horizon tool use and long-chain execution than Hunyuan 3. Text in / text out, 1M context, up to 64K output. Deep thinking is on by default and can be disabled; function calling, JSON mode, and implicit prompt caching supported.

StepFun Step 5 Preview, a 600B sparse-MoE (27B active) model built for agentic coding, software engineering and finance tasks. Text + image in / text out, 1M context. Thinking is always on and cannot be disabled; the trace is returned in reasoning_content and its tokens are billed inside completion_tokens, so max_tokens must cover thinking plus the answer. Function calling, JSON mode and implicit prompt caching supported.

Zhipu AI GLM-5.1, purpose-built for long-horizon tasks; 744B foundation, 200K context.

Qwen3.7-Max

Alibaba Qwen3.7-Max with native function calling and web search; 1M context.

Zhipu AI GLM-5V-Turbo, the vision-language turbo variant of GLM-5. Image and text in, text out, 200K-token context, deep-thinking mode, function calling, structured output, and implicit prompt caching.

Zhipu AI GLM-5.2-Fast-Preview, the low-latency preview build of GLM-5.2. Text in / text out, 1M-token context, hybrid thinking/non-thinking modes, function calling, structured output, and implicit prompt caching.

Qwen3.8-Max

Alibaba Qwen3.8-Max, a 2.4T-parameter MoE model for coding and agentic work; native visual understanding of documents and video, function calling and web search, 1M context.

Moonshot AI Kimi-K2.7-Code-HighSpeed, the high-throughput variant of Kimi-K2.7-Code. Image and text in, text out, 256K-token context, deep thinking, function calling, and implicit prompt caching.

Qwen3.8-2.4T-A95B

Alibaba Qwen3.8-2.4T-A95B, the 2.4T-parameter / 95B-active MoE member of the Qwen3.8 series. Text in / text out, native 1M-token context, thinking on by default, function calling, structured output, and implicit prompt caching.

Moonshot AI Kimi K3; text and image input, text output. 1M context, deep thinking on by default and can be disabled. Function calling and implicit prompt caching supported.

ByteDance Seedance 2.0

ByteDance Seedance 2.0 text-to-video and image-to-video, with first/last-frame control and multimodal references (images, video, audio). 4–15 second clips at 480p, 720p, 1080p or 4K, with optional generated audio. Billed on the tokens the vendor reports for each clip; jobs with a reference video use a lower rate.

ByteDance Seedance 2.5

ByteDance Seedance 2.5 text-to-video and image-to-video, with first/last-frame control and multimodal references (images, video, audio). 4–30 second clips at 480p, 720p or 1080p, with optional generated audio. Billed on the tokens the vendor reports for each clip; jobs with a reference video use a lower rate.

MiniMax-H3

MiniMax-H3 (Hailuo-03) text-to-video and image-to-video, with first/last-frame control and multimodal references (images, video, audio). 4–15 second clips at 2K. Billed per second of output, plus per second of reference video; reference images beyond five are charged per image.

Qwen2.5-VL-72B-Instruct

Alibaba Qwen2.5-VL 72B Instruct, a vision-language model for visual reasoning, OCR and document understanding; text and image input, text output, 128K-token context.

The smallest tier of the OpenAI GPT-5 line, for high-volume, latency-sensitive work; text and image input, text output, 400K-token context, function calling and implicit prompt caching.

The smallest tier of OpenAI GPT-4.1, for high-volume classification, extraction and other light tasks; text and image input, text output, 1M-token context, function calling and implicit prompt caching.

Gemini 2.5 Flash-Lite

The cheapest tier of the Google Gemini 2.5 line, for classification, extraction and other light tasks; text and image input, text output, 1M-token context, function calling supported.

The lightweight build of Z.ai GLM-4.5, with fewer active parameters for lower cost and latency; text in, text out, 128K-token context, function calling and implicit prompt caching.

The smallest, lowest-latency tier of the OpenAI GPT-5.4 line; text and image input, text output, 400K-token context, function calling and implicit prompt caching.

Fast, cost-efficient model in the GPT-5.6 family, optimized for high-volume, cost-sensitive workloads: responsive chat, classification, extraction, lightweight coding, and agentic workflows at lower latency and cost. Text and image input, text output, 1M-token context.

The mid-size tier of the OpenAI GPT-5 line, for everyday tasks at lower cost than the full model; text and image input, text output, 400K-token context, function calling supported.

Gemini 3.1 Flash-Lite

The low-latency, low-cost tier of the Google Gemini 3.1 line, for high-volume tasks; text and image input, text output, 1M-token context, function calling supported.

DeepSeek-V3

671B mixture-of-experts model; strong at coding, math, and multi-step reasoning.

Gemini 3.5 Flash-Lite

The cost-efficient tier of the Google Gemini 3.5 line; text and image input, text output, 1M-token context, function calling and implicit prompt caching.

DeepSeek-V3.1

DeepSeek-V3.1, the hybrid-thinking successor to V3; text in, text out, 160K-token context, function calling supported.

Gemini 2.5 Flash

Google Gemini 2.5 Flash, the fast tier of the Gemini 2.5 line, for high-volume work; text and image input, text output, 1M-token context, function calling supported.

DeepSeek-V3.1-Terminus

The Terminus build of DeepSeek-V3.1, tuned for agentic and tool-use workloads; text in, text out, 160K-token context, function calling and implicit prompt caching.

The mid-size tier of OpenAI GPT-4.1, trading some capability for lower cost and latency; text and image input, text output, 1M-token context, function calling and implicit prompt caching.

DeepSeek-R1

DeepSeek-R1, a reasoning model that emits a chain of thought before its answer; text in, text out, 64K-token context, function calling and implicit prompt caching.

Gemini 3 Flash (preview)

Preview build of the fast tier of the Google Gemini 3 line; text and image input, text output, 1M-token context, function calling supported. Preview: behaviour may change without a new model id.

Z.ai GLM-4.5, a mixture-of-experts model for reasoning, coding and agent workflows; text in, text out, 128K-token context, function calling and implicit prompt caching.

The 2025-09-05 snapshot of Moonshot AI Kimi K2, a mixture-of-experts model tuned for agentic coding and tool use; text in, text out, 256K-token context, function calling and implicit prompt caching.

The extended-reasoning build of Moonshot AI Kimi K2, which thinks across long tool-calling chains; text in, text out, 256K-token context, function calling supported.

The mid-size tier of the OpenAI GPT-5.4 line; text and image input, text output, 400K-token context, function calling and implicit prompt caching.

Z.ai GLM-4.7 for reasoning, coding and agent workflows; text in, text out, 200K-token context, function calling and implicit prompt caching.

OpenAI GPT-5.1 for reasoning, coding and agentic workflows; text and image input, text output, 400K-token context, function calling and implicit prompt caching.

Gemini 2.5 Pro

The reasoning tier of the Google Gemini 2.5 line, for coding, math and long-document analysis; text and image input, text output, 1M-token context, function calling supported.

LongCat 2.0

Meituan LongCat 2.0, a long-context mixture-of-experts model; text in, text out, 1M-token context, function calling and implicit prompt caching.

Qwen3-Coder-Plus

Alibaba Qwen3-Coder-Plus, tuned for repository-scale code generation and agentic coding; text in, text out, 1M-token context, function calling and implicit prompt caching.

Gemini 3.6 Flash

The fast tier of the Google Gemini 3.6 line; text and image input, text output, 1M-token context, function calling supported.

Gemini 3.5 Flash

The fast tier of the Google Gemini 3.5 line; text and image input, text output, 1M-token context, function calling and implicit prompt caching.

OpenAI GPT-5.2 for reasoning, coding and agentic workflows; text and image input, text output, 400K-token context, function calling supported.

Qwen3-Max

The largest tier of the Alibaba Qwen3 line; text in, text out, 256K-token context, function calling and implicit prompt caching.

Anthropic Claude Sonnet 5; text and image input, text output, with a 1M-token context window. Extended thinking and tool use supported.

Qwen3.6-Max (preview)

Preview build of the largest tier of the Alibaba Qwen3.6 line; text in, text out, 256K-token context, function calling supported. Preview: behaviour may change without a new model id.

OpenAI GPT-6.1 Sol, the balanced tier of the GPT-6 line for complex coding, computer use, and professional work, combining long-context reasoning with vision and tool calling. Text and image input, text output, 1.05M-token context, configurable reasoning effort.

OpenAI GPT-4.1 for general chat, coding and instruction following; text and image input, text output, 1M-token context, function calling and implicit prompt caching.

Gemini 3.1 Pro (preview)

Preview build of the reasoning tier of the Google Gemini 3.1 line; text and image input, text output, 1M-token context, function calling supported. Preview: behaviour may change without a new model id.

Balanced model in the GPT-5.6 family, tuned for workloads that need strong reasoning, coding, and agentic capability at lower cost than the flagship tier. Text and image input, text output, 1M-token context.

The higher-capability tier of the OpenAI GPT-5.4 line; text and image input, text output, 1M-token context, function calling and implicit prompt caching.

Anthropic Claude Sonnet 4.5; text and image input, text output, 1M-token context. Extended thinking and tool use are supported. Served over the Anthropic Messages format.

Anthropic Claude Sonnet 4.6; text and image input, text output, 1M-token context. Extended thinking, tool use, and explicit cache_control prompt caching are supported. Served over the Anthropic Messages format.

Anthropic Claude Opus 5.5; text and image input, text output, 1M-token context and up to 128K output tokens. Adaptive thinking is always on, and tool use and explicit cache_control prompt caching are supported. Served over the Anthropic Messages format.

Anthropic Claude Opus 4.6; text and image input, text output, 1M-token context. Extended thinking, tool use, and explicit cache_control prompt caching are supported. Served over the Anthropic Messages format.

Anthropic Claude Opus 4.7; text and image input, text output, 1M-token context. Extended thinking, tool use, and explicit cache_control prompt caching are supported. Served over the Anthropic Messages format.

Anthropic Claude Opus 5; text and image input, text output, with a 1M-token context window. Extended thinking and tool use supported.

GPT-5.5, designed for complex professional workloads, with strong reasoning, high reliability, and improved token efficiency on hard tasks. Text and image input, text output, 1M-token context.

Flagship of the GPT-5.6 series, built for advanced reasoning, complex coding, and agentic workflows: multi-step software engineering, long-horizon problem solving, and autonomous tool use. Text and image input, text output, 1M-token context.

Anthropic Claude Opus 4.8; text and image input, text output, with a 1M-token context window. Extended thinking and tool use supported.

Anthropic Claude Fable 5.1; text and image input, text output, 1M-token context. Extended thinking, tool use, and explicit cache_control prompt caching are supported. Served over the Anthropic Messages format.

OpenAI GPT-6 Astra, an agentic model built for long end-to-end tasks: reasoning, software engineering, computer use, browsing, scientific analysis, research, and professional document creation in one model. Text and image input, text output, 1.05M-token context. Sampling controls are fixed upstream; only reasoning depth is tunable.

Anthropic Claude Fable 5; text and image input, text output, with a 1M-token context window. Extended thinking and tool use supported.

ByteDance Seedream 5.0 Flash

ByteDance Seedream 5.0 Flash text-to-image with reference images, billed per generated image; the fastest Seedream. One image per request (a larger `n` is refused with 400). `size` is `<width>x<height>`, each side up to 4096 and 921,600–4,624,220 pixels in total (e.g. 1024x1024, 2048x2048 or 4096x1024); omitted, the model picks a size to fit the prompt. Up to 10 reference images in `image` (a string or an array of publicly reachable HTTP or `data:image/...` URLs).

Tencent Hy-Image-3.5 Preview

Tencent Hunyuan Hy-Image-3.5 (preview) text-to-image with reference images and multi-turn editing, billed per generated image at the price of its size tier: 1K/2K (output area up to 2048x2048) or 4K (above). Synchronous: POST /v1/images/generations; one image per request (`n` must be 1). Up to 20 reference images in `image` or `images` (a string or an array of HTTP or `data:image/...` URLs). Multi-turn: send the previous response's `assembled_history` as `messages` with the new instruction as `prompt`, and the same `session` string on every turn; each response returns the next `assembled_history`. `size` is `<width>x<height>`, at most 4096x4096 in area (e.g. 2048x2048, or 4096x4096 at the 4K price); omitted, `generate_max_pixels` picks the area (1048576, 2359296 or 4194304; default 2359296). Optional: `seed` (integer), `resize_max_pixels` (area cap for reference images), `footnote` (watermark text, up to 16 characters), `use_search_tool` (true: let the model search the web).

Tencent Hy-Image-3.0

Tencent Hunyuan Hy-Image-3.0 text-to-image with reference images, billed per generated image. Synchronous: POST /v1/images/generations; one image per request (`n` must be 1). Up to 3 reference images in `image` or `images` (a string or an array of HTTP or `data:image/...` URLs). `size` is `<width>x<height>`, each side 512–2048 and at most 1024x1024 in area; omitted, the model picks one. Optional: `seed` (integer), `revise` (true/false: let the model rewrite the prompt), `footnote` (watermark text, up to 16 characters).

ByteDance Seedream 4.5

ByteDance Seedream 4.5 text-to-image with reference images, billed per generated image. One image per request (a larger `n` is refused with 400). `size` is `<width>x<height>` with each side up to 4096 and at least 3,686,400 pixels in total (e.g. 2048x2048 or 2560x1440; 1024x1024 is refused); omitted, the model renders 2048x2048. Up to 14 reference images in `image` (a string or an array of publicly reachable HTTP or `data:image/...` URLs).

ByteDance Seedream 5.0 Lite

ByteDance Seedream 5.0 Lite text-to-image with reference images, billed per generated image. One image per request (a larger `n` is refused with 400). `size` is `<width>x<height>` with each side up to 4096 and at least 3,686,400 pixels in total (e.g. 2048x2048 or 2560x1440; 1024x1024 is refused); omitted, the model renders 2048x2048. Up to 14 reference images in `image` (a string or an array of publicly reachable HTTP or `data:image/...` URLs).

MiniMax Image-01

MiniMax image-01 text-to-image, billed per generated image. Synchronous: POST /v1/images/generations; `n` up to 9.